DeepSeek-V4.1-Flash

使用前按你的设备修改模型路径和服务地址，并核对上方版本与参数。

sudo apt-get update
sudo apt-get install -y git curl rsync python3-venv rdma-core ibverbs-utils unzip
uname -m                 # 应为 aarch64
nvidia-smi               # 驱动正常识别计算设备
docker version
ip -br addr
rdma link show
ibv_devinfo
# 如系统已提供 show_gids，可据此选择 RoCE v2 IPv4 GID。

mkdir -p "$HOME/megacube-deepseek"
unzip megacube-deepseek-deploy.zip -d "$HOME/megacube-deepseek"

python3 -m venv "$HOME/.venvs/model-download"
source "$HOME/.venvs/model-download/bin/activate"
pip install modelscope
mkdir -p "$HOME/models/DeepSeek-V4.1-Flash"
modelscope download --model deepseek-ai/DeepSeek-V4.1-Flash \
  --local_dir "$HOME/models/DeepSeek-V4.1-Flash"

# 在解压的部署包目录执行；校验索引、分片长度及 Engram 容量。
python3 "$HOME/megacube-deepseek/verify-model.py" "$HOME/models/DeepSeek-V4.1-Flash"
cd "$HOME/models/DeepSeek-V4.1-Flash"
find . -type f ! -name SHA256SUMS -print0 | sort -z | \
  xargs -0 sha256sum > SHA256SUMS
for host in cube-b cube-c cube-d; do
  ssh "$host" 'mkdir -p "$HOME/models/DeepSeek-V4.1-Flash"'
  rsync -a --partial --append-verify "$HOME/models/DeepSeek-V4.1-Flash/" \
    "$host:models/DeepSeek-V4.1-Flash/"
  ssh "$host" 'cd "$HOME/models/DeepSeek-V4.1-Flash" && sha256sum -c SHA256SUMS'
done

cd "$HOME/megacube-deepseek"
bash build-image.sh

cd "$HOME/megacube-deepseek"
for host in cube-b cube-c cube-d; do
  ssh "$host" 'mkdir -p "$HOME/megacube-deepseek"'
  rsync -a --exclude='.git' spark-vllm "$host:megacube-deepseek/"
  rsync -a start-rank.sh cluster.env.example "$host:megacube-deepseek/"
  docker save megacube/deepseek-v41:release-20260913 | ssh "$host" 'docker load'
done
# 每台执行一次；应正常加载预编译的 MXFP8 和 sparse MLA 内核。
docker run --rm --gpus all --entrypoint python3 \
  megacube/deepseek-v41:release-20260913 /opt/dsv41-verify-kernels.py

# cube-d
cd "$HOME/megacube-deepseek" && bash start-rank.sh 3
# cube-c
cd "$HOME/megacube-deepseek" && bash start-rank.sh 2
# cube-b
cd "$HOME/megacube-deepseek" && bash start-rank.sh 1
# 最后在 cube-a
cd "$HOME/megacube-deepseek" && bash start-rank.sh 0
docker logs -f megacube-deepseek-v41

curl -f http://127.0.0.1:18411/health
curl -f http://127.0.0.1:18411/v1/models
curl -N http://127.0.0.1:18411/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model":"deepseek-v4.1-flash","messages":[{"role":"user","content":"请用 Python 实现合并重叠区间的函数。"}],"temperature":0,"max_tokens":256,"stream":true,"chat_template_kwargs":{"thinking":false}}'

cd "$HOME/megacube-deepseek"
python3 spark-vllm/tools/vision_tools_demo.py http://127.0.0.1:18411/v1
mkdir -p results
python3 spark-vllm/bench/v41bench.py \
  --base http://127.0.0.1:18411/v1 --model deepseek-v4.1-flash \
  --label megacube-local --out results